Struct AgentConfig 

pub struct AgentConfig {

    pub path: TaskPath,
    pub base_system_prompt: Option<String>,
    pub prompt_assembly: Option<Arc<PromptAssembly>>,
    pub prompt_profile: PromptProfile,
    pub max_iterations: Option<usize>,
    pub cwd: PathBuf,
    pub env: HashMap<String, String>,
    pub max_tool_result_chars: usize,
    pub max_message_tool_results_chars: usize,
    pub policies: Arc<PolicyRegistry>,
    pub storage: Option<Arc<dyn Storage>>,
    pub tool_diagnostics: Option<Arc<dyn ToolDiagnosticsSink>>,
    pub compaction: Option<Arc<dyn HistoryCompactionStrategy>>,
    pub permission_engine: Option<PermissionEngine>,
    pub approval_handler: Option<Arc<dyn ApprovalHandler>>,
    pub tool_concurrency_limit: usize,
    pub token_budget: Option<TokenBudget>,
    pub retry: RetryConfig,
    pub auto_validate_schema: bool,
    pub cancellation: CancellationState,
    pub tool_timeout_ms: Option<u64>,
    pub max_file_read_bytes: usize,
    pub skill_registry: Option<Arc<SkillRegistry>>,
    pub plugin_registry: Option<Arc<PluginRegistry>>,
    pub task_manager: Option<Arc<TaskManager>>,
    pub memory_injector: Option<Arc<MemoryInjector>>,
    pub skill_injector: Option<Arc<SkillInjector>>,
    pub mcp_manager: Option<Arc<McpManager>>,
    pub event_channel: Option<EventChannelConfig>,
}

Configuration for an AgentSession.

Build with Default::default and override fields as needed. All fields are public so callers don’t need a builder for simple cases.

Fields

path: TaskPath

Workflow path classification for the current task. Influences iteration budget, timeout, concurrency, and the system-prompt guidance injected at turn time.

base_system_prompt: Option<String>

Optional base instruction appended to the identity section of the system prompt. For full control, use prompt_assembly instead.

prompt_assembly: Option<Arc<PromptAssembly>>

Optional pre-built prompt assembly. When set, the runtime uses this instead of constructing the prompt from base_system_prompt alone.

prompt_profile: PromptProfile

Controls how much built-in prompt guidance is injected when the runtime constructs the default prompt assembly.

max_iterations: Option<usize>

Optional maximum number of model ⇄ tool round-trips per turn before the loop aborts with AgentError::MaxIterations.

None means the model/tool loop is allowed to continue until the model finishes, the turn is cancelled, or another runtime limit is reached.

cwd: PathBuf

Working directory used as the root for filesystem tools and shell commands.

env: HashMap<String, String>

Environment variables exposed to shell-based tools.

max_tool_result_chars: usize

Hard cap on the character length of any individual tool result. Anything longer is replaced with a truncated preview to protect the context window.

max_message_tool_results_chars: usize

Hard cap on the aggregate character length of all tool results within a single message. When N parallel tool calls collectively exceed this, the largest results are truncated to fit. Prevents a turn with many tool calls from flooding the context window even when each result individually stays under max_tool_result_chars.

policies: Arc<PolicyRegistry>

Registry of semantic runtime policies.

storage: Option<Arc<dyn Storage>>

Optional persistent backing store for session messages.

tool_diagnostics: Option<Arc<dyn ToolDiagnosticsSink>>

Optional sink for sanitized tool failure diagnostics.

compaction: Option<Arc<dyn HistoryCompactionStrategy>>

Optional history-level compaction strategy (summarisation, etc.).

permission_engine: Option<PermissionEngine>

Optional rule-based permission engine consulted before every tool call.

approval_handler: Option<Arc<dyn ApprovalHandler>>

Optional handler invoked when a tool call requires explicit human approval.

tool_concurrency_limit: usize

Maximum number of concurrency-safe tools to run in parallel within a single batch.

token_budget: Option<TokenBudget>

Optional token budget that triggers proactive compaction.

retry: RetryConfig

Retry configuration for transient provider failures. When None, provider calls are not retried.

auto_validate_schema: bool

Whether to automatically validate tool arguments against the tool’s input_schema after the tool’s own validate method runs.

cancellation: CancellationState

Shared cancellation state. Hosts call cancel() to interrupt a running turn and reset() before starting the next turn.

tool_timeout_ms: Option<u64>

Per-tool execution timeout in milliseconds. When set, any tool that runs longer than this will be interrupted and its result will be an is_error: true timeout so the model can recover.

max_file_read_bytes: usize

Maximum number of bytes the built-in file tools will read from a single file. Files larger than this are rejected to avoid OOMing the agent.

skill_registry: Option<Arc<SkillRegistry>>

Optional skill registry. When set, the Skill tool and the default prompt assembly can use registered skills (including bundled Superpowers skills).

plugin_registry: Option<Arc<PluginRegistry>>

Optional plugin registry. When set, enabled plugins’ components are applied to tool/policy/skill registries at session startup.

task_manager: Option<Arc<TaskManager>>

Optional task manager used by long-running tools such as background subagents to publish lifecycle state and output.

memory_injector: Option<Arc<MemoryInjector>>

Optional memory injector for dynamic per-turn memory scoring. When set, memories are scored against the user’s current input and injected as system reminders before the first provider call of each turn.

skill_injector: Option<Arc<SkillInjector>>

Optional skill injector for dynamic per-turn skill discovery. When set, top matching skills are injected as system reminders before the first provider call of each turn.

mcp_manager: Option<Arc<McpManager>>

Optional MCP manager. When set, MCP server tools are bridged into the tool registry and an MCP tools section is added to the prompt assembly.

event_channel: Option<EventChannelConfig>

Optional configuration for the bidirectional HTTP event channel. When set and enabled is true, the session starts an embedded HTTP server on the configured address, accepting external event injection (POST /inject) and streaming TurnEvents via SSE (GET /events).

Implementations

impl AgentConfig

pub fn request_cancel(&self)

Request cancellation and wake any runtime await that is listening for it.

pub fn reset_cancel(&self)

Clear a previous cancellation request before starting a new turn.

pub fn validate(&self) -> Result<(), AgentError>

Validate configuration values and return a structured error for any dangerous or nonsensical combination.

pub fn with_default_prompt_assembly( self, tools: Arc<ToolRegistry>, ) -> Result<Self, AgentError>

Replace base_system_prompt with the default modular prompt assembly.

This is the recommended way to use telos-agent’s built-in default prompt sections without manually wiring each section.

pub fn with_path(self, path: TaskPath) -> Self

Apply path-appropriate defaults for timeouts, token budgets, and concurrency. Call after setting custom values if you want the path to override them.

Knob	Fast	Standard	Heavy
max_iterations	None	None	None
tool_concurrency_limit	10	10	5
token_budget	None	None	308k
tool_timeout_ms	30_000	None	60_000

pub fn with_bundled_skills(self) -> Self

Load bundled skills into a fresh skill registry attached to this config.

pub fn apply_plugins( &self, tools: ToolRegistry, policies: PolicyRegistry, skills: SkillRegistry, mcp: McpManager, prompt: PromptAssembly, ) -> (ToolRegistry, PolicyRegistry, SkillRegistry, McpManager, PromptAssembly, Result<(), Vec<PluginError>>)

Apply plugin components into the agent registries.

Call this before creating an AgentSession to populate tools/policies/skills/mcp/prompt with enabled plugin content. Always returns the registries (even on error) so callers can continue with degraded state.

Trait Implementations

impl Clone for AgentConfig

fn clone(&self) -> AgentConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for AgentConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for AgentConfig

fn default() -> Self

Returns the “default value” for a type.